Red teaming
Also known as: red team · red-teaming an LLM
Red teaming is deliberately trying to break your AI system — adversarial prompts, jailbreaks, edge cases — to find weaknesses before an attacker does.
- Manual: a human (or another LLM) probes the model.
- Automated: batches of adversarial prompts.
- Output: a list of failure modes and fixes.
The mindset: red teaming assumes you *will* be attacked. Finding your own jailbreaks on day one is far cheaper than a customer finding them on day one hundred.